ESTRO 2024 - Abstract Book

S3906

Physics - Image acquisition and processing

ESTRO 2024

While diffusion models are capable of generating high-quality images, they also introduce diversity, resulting in structural changes in the output images. This aspect of the diffusion model, or more broadly, generative AI, is considered undesirable in medical applications, particularly in image guidance for IGRT. To mitigate these structural changes, we combine the diffusion model with an iterative CT reconstruction technique commonly employed to enhance the image quality of CBCT and MVCT. Since the projection data used in CT reconstruction contains information about the patient's anatomical structure, enforcing a consistency loss between the projection data and the output from the diffusion model allows us to preserve these structures in the images without compromising the overall image quality. We utilize the Denoising Diffusion Probabilistic Model (DDPM) and train it with a substantial volume of high-quality images. Subsequently, using the trained denoising model, we reconstruct CT images by optimizing the latent variable while keeping the model parameters fixed to minimize the consistency loss. The schema of the proposed method is shown in Figure 1.

Results: